Synthetic data issues

[qqhe-frank]

Very nice work, I would like to follow the process provided by the authors in the paper to generate synthetic data for training, I have a query for the authors, for the synthetic data, the HR image is offset and then downsampled, does this make the estimated motion offset also scaled down?
For example, if I offset the HR image by 0.5 pixels and then downsample it by 2x to get the LR image, and then estimate the motion of the LR image against the reference image, will the estimated motion be 0.25 pixel values?